A Heterogeneous Ensemble Learning Method Combining Spectral, Terrain, and Texture Features for Landslide Mapping

Abstract

The existing landslide recognition methods mainly focus on the use of spectral bands of optical remote sensing and machine learning base classifiers, which are insufficient in landslide characterization in complex scenes, resulting in a high missed and false detection of landslides. In this article, we develop a landslide recognition framework, which combines the multidimensional feature advantages of spectral, terrain, and texture of optical satellite images, and constructs a heterogeneous ensemble learning method for landslide mapping. First, we construct a landslide multidimensional feature dataset using Sentinel-2A and Advanced Land Observing Satellite digital elevation model data. Then, we construct a heterogeneous ensemble learning landslide recognition method, which combines the advantages of fully convolutional network, U-Net, and attention U-Net base classifiers to fully learn the multidimensional features of landslides. Finally, we evaluate the performance of the landslide recognition framework in the Bailongjiang River Basin complex scenes. The experimental results show that integrating the multidimensional features of spectral, terrain, and texture and using the heterogeneous ensemble learning method can reduce the missed and false detection of landslides in complex scenes. Specifically, compared with using only spectral bands, integrating spectral bands, spectral indexes, terrain factors, and texture indexes achieves the highest Recall, Kappa, F1-score, and MIoU in testing areas, and missed alarm (MA) is reduced by 15.56%. Compared with deep learning base classifiers, the constructed heterogeneous ensemble learning demonstrates improvements in Recall ranging from 41.67% to 69.89%, and MA is reduced from 52.17% to 30.11%. This study provides a new idea for high-precision landslide recognition in complex environments.